Perceptual analysis of higher-order statistics in estimating reverberation

This paper presents a study of the capacity of four speech signal features to assess speech perceptual quality and their use in a typical two-stage algorithm for reverberant speech enhancement. This algorithm is divided into two blocks: one that deals with the coloration effect, due to the early reflections, and the other for reducing the long-term reverberation. The proposed features are skewness, two types of kurtosis and Shannon´s entropy. This assessment capacity is evaluated by two perceptual-quality measure specific for the speech-reverberation context. Experimental results for a 204-signal database show that the proposed features can achieve a correlation coefficient of -75% (for entropy) which indicates the potential use for entropy in speech enhancement algorithms.